When Alpha Breaks: Two-Level Uncertainty for Safe Deployment of Cross-Sectional Stock Rankers

Battle Count: 82

Directly addresses the core operational challenge of deploying ML-based cross-sectional stock rankers under non-stationarity. Provides a complete, implementable framework for when to trade (regime gate), how to size positions (volatility sizing), and how to manage tail risk (epistemic cap). The structural coupling finding (ρ(ê,|score|)=0.616) is a critical insight for any quant deploying uncertainty-aware ranking systems. The paper demonstrates that standard inverse-uncertainty sizing from the return-prediction literature does NOT transfer to ranking, which is a fundamental practical finding. Multi-crisis validation across 8 windows and cross-model robustness testing strengthen deployment confidence.

The architecture is operationally simple (binary gate + volatility sizing + percentile cap) but requires careful PIT-safe implementation. Key complexity sources: (1) walk-forward training with 90-day embargo and label purging, (2) PIT-safe aleatoric baseline construction from matured losses, (3) regime-trust gate requiring EWMA of matured RankIC with structural lag plus drift/disagreement components, (4) conformal prediction calibration with rolling windows. The base LightGBM ranker uses only 7 features. The DEUP error predictor uses 11 features with conservative hyperparameters. All components are computationally lightweight (tree-based models). The main implementation challenge is ensuring strict point-in-time safety throughout the pipeline.

GitHub repository is referenced in the paper. Walk-forward protocol with 109 folds, DEV/FINAL split, and PIT-safe construction are clearly described. Hyperparameters for LightGBM base model and DEUP error predictor are specified. However, the specific AI-themed ticker list (100 names) and vendor data endpoints are not fully enumerated. The paper provides detailed feature lists, training protocols, and evaluation metrics. All deployment-facing results use PIT-safe aleatoric baselines with explicit ablation reporting.